Single-image 3D scene completion without the LiDAR bill

What it does MonoScene takes a single RGB image and outputs a 3D semantic occupancy grid — predicting both visible and occluded geometry along with per-voxel class labels. It targets outdoor driving data (SemanticKITTI, KITTI-360) and indoor scenes (NYUv2), offering pretrained checkpoints for immediate evaluation.

The interesting bit The repository treats monocular scene completion as a practical prediction problem rather than a theoretical exercise, shipping with a Hugging Face demo and multi-GPU training scripts. That said, the README is almost entirely a setup manual; the actual network architecture and loss design are left to the paper.

Key highlights

• Single-image inference on SemanticKITTI, NYUv2, and KITTI-360
• Pretrained models available for direct download and evaluation
• Live demo hosted on Hugging Face for quick browser testing
• Visualization pipeline via mayavi, with an Open3D workaround noted in the issues
• Multi-GPU PyTorch training scripts included for SemanticKITTI and NYUv2

Caveats

• The README is a setup guide with almost no architectural detail; you will need the paper to understand how the model actually works.
• Dependency stack is locked to older releases (Python 3.7, PyTorch 1.7.1, CUDA 10.2).
• KITTI-360 support is inference-only; training is limited to SemanticKITTI and NYUv2.

Verdict A solid CVPR 2022 baseline if you are benchmarking camera-only 3D occupancy methods. Look elsewhere if you need a modern, well-documented training framework or turnkey dependency management.